Workshop Digital Twins for Pipe Transport Networks

Claire Heaney
Imperial College London, Faculty of Engineering, Department of Earth Science & Engineering, London, United Kingdom

AI-based Reduced-Order Modelling of Pipe Flows

This presentation will discuss different aspects of modelling two-phase flows in pipes exploiting Artificial Intelligence (AI) technologies. We present an AI-based data-driven or non-intrusive reduced-order model that can make predictions for unseen scenarios with larger-sized domains than were used in training. The approach sets the reduced-order model within a domain decomposition framework, exploits a sub-sampling technique to obtain snapshots from computational fluid dynamics simulations and uses AI methods. To obtain a low-dimensional space for the reduced-order model, a convolutional autoencoder is used, as this type of network can compress information more efficiently than traditional approaches. For prediction, an adversarial network is used, which attempts to keep the predictions realistic. The method is applied to chaotic time-dependent multiphase flow in a pipe. The approach presented here shows great potential for increasing the generalisation of data-driven reduced-order models. More details about the methods and results described in this presentation can be found in.

Sridharakumar Narasimhan
Indian Institute of Science Madras, Department of Chemical Engineering, Chennai, India

Title: TBA

Zoran Kapelan
Delft University of Technology, Department Water Management, Faculty of Civil Engineering and Geosciences, Delft, The Netherlands

AI-based Detection of Pipe Bursts/Leaks

The keynote will start with introductory notes about the water leakage in drinking water distribution systems and the related summary of principal existing approaches for the detection of pipe bursts/leaks in these systems. The main part of the lecture will present an AI-based methodology for the automated detection and location of pipe bursts/leaks but also equipment and other failures in drinking water systems. This technology detects pipe bursts/leaks by processing pressure, flow and other sensor data in near real-time by using different machine learning type methods. Elements of this methodology were built into a commercial solution that has been used companywide in a large water utility resulting in major operational cost savings. Other examples of AI-based methods for the detection of other events such as water quality issues in drinking water systems, blockages in urban drainage pipes and failure events at water treatment plants will be presented briefly too. Most of these have been developed in collaboration with various water utilities and are based on real data and case studies. The talk will end with several take home messages for the researchers working in this field.

Mark Roest
VORtech BV, Delft, The Netherlands

Digital twins and experiments in leak localization

Mark Roest is senior partner at VORtech, a company of computational software engineers. VORtech works for a broad range of large companies and institutes to develop and improve their computational software. The concept of digital twins is interesting for many of VORtechs clients. Mark will briefly describe several aspects around digital twins for operational purposes.
Much of the work that VORtech has done over the past decades can, in hindsight, be qualified as related to the digital twin concept. One example is the work that was done on leak localization, first by order of Vitens directly and later in collaboration with Deltares. The initial idea was to find leakages by calibrating a transport network model with unknown leakages, based on observations from the real network. This was not immediately successful.
Then, a very brute force approach was adopted to just simulate many leakage scenarios and match the simulation results with the observed values from the real network. A relatively simple but effective search algorithm proved to be sufficient to find many leakages. Further experiments have been done to assess the conditions under which this approach does or does not work.

Prerna Pandey
Indian Institute of Science Bangalore, Management Studies, Bangalore, India

A Two-Stage Model for Data-Driven Leakage Detection and Localization in Water Distribution Networks

Water utilities face the challenge of reducing water losses by promptly detecting, localizing, and repairing leaks during their operational stage. To address this challenge, utilities are exploring alternative approaches to detect leaks with high accuracy in a timely manner, while minimizing environmental and economic consequences. In this study, we propose a data-driven two-stage model that leverages pressure and flow rate data monitored at several locations in a water distribution network (WDN) to predict the occurrence and location of leaks. The first stage involves computing the errors in predicting pressure heads, while the second stage uses statistical measures in the error distributions to classify the leak occurrence and location. The proposed approach is cost-effective and can be readily deployed.
Simulation-based case studies are used to demonstrate the effectiveness of the proposed model on several benchmark networks. Additionally, the methodology has been implemented in a real-life WDN network where sensors are situated at various points throughout the system, and real-time monitoring of flow and pressure data is carried out using an Internet of Things (IoT) system. The results of this approach indicate that the model accurately predicts the occurrence and location of leaks. However, the model's performance decreases as the network size increases, leading to reduced accuracy.

Nikolaj Mücke
Centrum Wiskunde & Informatica, Scientific computing, Amsterdam, The Netherlands

Probabilistic Digital Twins for Leak Localization in Pipe Flows

Digital Twins have recently emerged as a promising technology for predictive maintenance
and anomaly detection in complex systems. They work by creating a virtual replica of a
physical system, allowing for real-time monitoring and simulation.
In this talk, we present methodologies for digital twins that not only deliver predictions, but
also quantifies the uncertainty. This is essential in many modern industrial applications where
risk analysis has become an integral part of the decision making process.
Our methodologies are based on Bayesian inference and generative deep learning for
combining model simulations with sensors observations. Bayesian inference serves as the
framework for computing the posterior distribution over quantities of interest conditioned on
observations. However, Bayesian inference is typically too slow for real-time inference.
Therefore, we make use of generative deep learning as a stochastic surrogate model to speed
up the computations.
In this talk, we present two different leak localization problems. Firstly, we consider large
scale water distribution networks based modelled by steady state balance equations.
Secondly, we look at long single pipes modelled by dynamic partial differential equations.